Authority & Mentions · 9 min read · July 15, 2026
Documenting reference projects so AI treats them as evidence, not just photos
When someone asks ChatGPT which joiner near them builds a solid-wood kitchen or a made-to-measure fitted unit, the answer doesn't come from a hunch, it comes from what the model can actually read about your work. Reference projects are your best evidence, but only when they're written up in a way a machine can parse: material, dimensions, location, task, and result, spelled out in text. Here's how to do that.
Why your best photos are invisible to the AI
Most joiners have a gallery on their site: a couple dozen photos of kitchens, staircases, and fitted wardrobes, often without a single line of text. To a person, that's impressive. To ChatGPT, Perplexity or Google's AI, it's close to invisible. Language models read images on ordinary web pages only in a limited way, and their answers come overwhelmingly from text. A photo of a walnut kitchen island tells the model nothing if nowhere on the page does it say it's a walnut kitchen island.
That's the core mismatch in the trades: we show our craft visually, because that's how we experience it. But the system advising your next customer reads. It's looking for sentences like "built from oiled wild oak, 3.2 meters of worktop, installed in a period building in Regensburg." Without sentences like that, your best work stays a picture with no meaning attached.
The good news is you don't need to build anything new. You just need to describe what you've already built. And because almost none of your competitors do this, even a single cleanly documented project can put you ahead in AI-generated answers.
What turns a reference project into evidence
Proof isn't the same as praise. "We build high-quality furniture" is a claim. "In 2025 we built a dental practice in Augsburg a reception counter in oiled oak veneer with integrated cable management and LED lighting, 4.1 meters long, installed in a single day" is proof. The difference is verifiability: concrete numbers, materials, locations, and tasks read nothing like marketing copy.
Language models favor exactly this kind of verifiable statement, because it lets them answer a specific question. Ask "who builds solid-wood practice fit-outs in Swabia?" and the model can point to your documented counter project, name you, and confirm the material and the region. Without those details, you're just another anonymous joinery.
The rule of thumb: every sentence should carry a detail a customer couldn't have made up. Material, wood species, dimensions, the installation situation, the tricky part, the timeframe. The more specific it is, the more it's believed, by people and by the model.
The five parts every project write-up needs
Use a fixed structure so you're not reinventing it for every project. Five parts cover it. First, the task: what did the customer need, and what was the problem? Second, the material: wood species, surface finish, hardware, board material. Third, the dimensions and key numbers: length, height, the room, the quantity. Fourth, the complication: what was slanted, awkward, under heritage protection, or otherwise unusual? Fifth, the result: what did the customer end up with, and how long did it take?
For a kitchen, that might read: "Task: open-plan kitchen in a terraced house in Fürth; the customer wanted seamless fronts and maximum storage on a tight footprint. Material: handleless MDF fronts in matte anthracite, worktop in 4-centimeter solid oak, soft-close pull-outs. Dimensions: 3.8 meters of run plus a 2.1-meter island. Complication: a sloped ceiling above the sink, built entirely to fit. Result: handed over ready to use, six weeks from order."
You can fill in these five parts in a few minutes per project, ideally right after installation while the details are still fresh. Keep a simple template on your phone and dictate it on the drive home from the job. That builds a library of real, verifiable references with almost no extra effort.
Captions and alt text: the lever nobody pulls
Your photos still matter, for people. But you can make them work for the AI too, by giving every image a real alt text and a real caption. Alt text is the description behind the scenes that search engines and AI systems actually read. Instead of "IMG_4821.jpg" or "kitchen," write something like: "kitchen island in oiled wild oak with a downdraft extractor, built by our joinery in Regensburg."
Every project photo should also carry a visible caption, written as a full sentence. "Fitted wardrobe in an attic, sloped panels in birch plywood, painted white, 3.4 meters wide" is worth far more than no caption at all. These small blocks of text add up: a gallery of twenty well-captioned photos gives the AI twenty usable pieces of evidence instead of twenty silent files.
Keep the descriptions honest and specific. No keyword stuffing, no invented superlatives. Both the AI and attentive customers spot filler. A plain, precise sentence about what's actually in the photo reads as the most credible.
Getting customer reviews to actually help
Reviews are a strong signal for language models, because they're independent confirmation. But a star rating with no text does little for the AI. So ask customers directly for a few sentences naming the actual project: "The joinery built our living-room wall completely to measure, with a built-in shelf in ash and a cutout for the fireplace." Specific reviews like that tie your name to a real, describable piece of work.
It helps most when the same details show up in more than one place: your project write-up, the customer's Google review, maybe an industry directory. When an AI model sees the same fact confirmed from multiple independent sources, it trusts that detail more. Contradictions do the opposite, so keep your business name, location, and services worded identically everywhere they appear.
Ask right after handover, while the customer is still happy with the finished piece. A short line like "if you're happy with it, a few sentences on Google about what we built for you would help us a lot" gets you far better reviews than a bare star rating weeks later.
Give each project its own page
The strongest move is a dedicated page for every larger project, instead of burying it in a crowded gallery. A page titled "Solid-wood open kitchen in Fürth with seamless anthracite fronts" gives the AI something it can clearly tie to a topic, a location, and a service. That specificity is exactly what generative systems are looking for when they answer a location-specific question.
On that page, work in all five parts, add two or three captioned photos, and include the customer's own words about the project if you have them. That gives you one closed loop of proof: task, execution, material, result, and confirmation, all in one place. Ten pages like that do more for your AI visibility than a gallery of a hundred photos with no captions at all.
Think in your customers' actual questions. Nobody searches "furniture making"; they search "joiner for a walk-in closet under a sloped roof" or "bakery shop fitting in oak." When one of your project pages matches that exact situation, you become the answer. Write your page titles and opening sentences in your customers' words, not your industry's.
Mistakes that quietly undercut your references
The most common mistake is generalizing. "We build custom furniture to the highest standard" is on every second joinery site, and it's worthless to an AI model because there's nothing in it to verify. Replace every line like that with a specific project and real numbers. One proven sentence beats ten pieces of ad copy.
The second mistake is contradictions and stale information. If your homepage says "since 1998," your legal notice gives a different founding year, and Google still has your old address, that inconsistency makes an AI model less confident in anything it finds about you. Keep your name, address, phone number, and service description identical everywhere they appear. It's unglamorous, but it matters.
The third mistake is leaving out the date entirely. A project with no year attached reads as old, whether it was or not. Date every reference project. "Completed 2025" tells both the AI and the customer that you're active and still doing exactly this kind of work now.
Where to start this week
You don't have to do this for everything at once. Pick your three best completed projects from the last two years and write down the five parts for each. That's roughly ten minutes per project, and it immediately gives you three real, verifiable reference texts that say more than your whole existing gallery.
Then make it a habit: at every handover, dictate the five parts into your phone on site, and ask the customer for a short, project-specific review. Your reference collection grows on its own after that, instead of becoming a scramble at the end of the year. Every finished job turns into another piece of your AI visibility.
The effect isn't instant, but it compounds. Ads only work while you're paying for them; well-documented references keep working indefinitely. Every project you describe stays a piece of evidence the AI can pull up again and again, the next time a customer asks for exactly the kind of joinery work you do.
Common questions
Isn't it enough to just upload a lot of photos of my work?
No. Photos land with people, but AI systems like ChatGPT or Perplexity draw their answers almost entirely from text. A photo of your oak kitchen stays invisible to the model unless the words next to it say it's an oak kitchen, with material, dimensions, and location spelled out. Add a specific caption and real alt text to every photo, and a silent picture becomes usable evidence.
Do I need to list prices on my reference projects?
No, that's not necessary, and with custom joinery work it's often not even sensible, since every piece is made to measure. Material, dimensions, task, complication, and timeframe matter more than the price. If you want to give a rough sense of scale, a range like "fitted wardrobes from around 2,500 euros" works. For AI visibility, the concrete, checkable details of the work count for more than the invoice total.
How many reference projects should I document?
Quality beats volume. Ten well-described projects, each with its own page, real material details, and a customer quote, carry more weight than a much longer list of thin gallery entries. Start with your three best jobs from the last two years, then add a new one at every handover. Aim for variety, kitchens, wardrobes, staircases, shop fitting, so you cover different customer questions and become the right answer for more of them.
Read on